Natural disasters and human migration in the United States: Insights from automated machine learning and explainable AI.

The growing threats of natural disasters influence human migration at various spatio-temporal scales. This study examines how the four major disaster types (i.e., floods, hurricanes, wildfires, and tornadoes) relate to human migration in the contiguous United States (CONUS) from 2000 to 2020. Utilizing statistical and machine learning methods, we quantify spatial and temporal variations in migration patterns in relation to disaster impacts while controlling for socio-economic and environmental variables. Results indicate that counties experiencing higher disaster impacts consistently show lower average net migration rates (NMR), with increased disaster frequency or damage often correlating with reduced migration rates. Using an automated machine learning (AutoML) framework, we developed predictive models that can explain 59% to 72% of the variance in county-level NMR, significantly outperforming benchmark linear regression models. SHapley Additive exPlanations (SHAP) values were used to assess the contribution of disaster and socio-economic variables to model predictions. Although socio-economic factors remain the dominant predictors, hurricanes, floods, and wildfires showed substantial associations with migration patterns over the two decades. Overall, this study demonstrates the value of explainable AI in capturing the complex dynamics between natural disasters and human migration, offering insights into how disasters and socio-economic factors are jointly associated with population movement.